Competition and Technology Adoption∗

Jeffrey Macher†

Georgetown UniversityNathan H. Miller‡

Georgetown University

Matthew Osborne§

University of Toronto

August 20, 2015

Abstract

We examine the adoption of precalciner technology in the portland cement industryover 1974-2010. Our estimation results indicate that adoption is positively affectedby the “fuel cost gap” between incumbent kilns and precalciner kilns, but that thisrelationship dissipates with both domestic and import competition. We motive thisinteractive effect of competition using a theoretical model adapted from Dasgupta andStigliz (1980). Interpreted through the model, our results indicate that competitionlimits the ability of firms to appropriate the benefits of new technology adoption. Ourresearch contributes to the literature on competition and innovation, and has practicalimplications for policy-makers.

Keywords: technology, innovation, competition, portland cementJEL classification: L1, L5, L6

∗We thank Erika Lim and Jiayi Zhang for excellent research assistance. We appreciate the helpfulcomments received at the Business and Public Policy Conference at ESMT in Berlin, Germany. The viewsexpressed herein are entirely those of the authors and should not be purported to reflect those of the U.S.Department of Justice.†Georgetown University, McDonough School of Business, 37th and O Streets NW, Washington DC 20057.

Email: jeffrey.macher@georgetown.edu.‡Georgetown University, McDonough School of Business, 37th and O Streets NW, Washington DC 20057.

Email: nhm27@georgetown.edu.§University of Toronto, Rotman School of Management, 105 St. George St., Toronto, ON, Canada M5S

3E6. Email: matthew.osborne@rotman.utoronto.ca

1 Introduction

Does competition spur innovation and new technology adoption? This question has occupied

economic researchers since Joseph Schumpeter (1942) posited that the stability of large

firms in concentrated markets gives rise to greater R&D investment.1 An early empirical

literature shows some propensity of firms in moderately concentrated industries to invest

more in R&D, but these effects largely dissipate in econometric studies that control for

industry-specific effects (e.g., Scott (1984); Levin et al (1985)). This development accords

with the theoretical literature: while market power can induce investment (e.g., Dasgupta

and Stiglitz (1980)), the opposite effect can arise if, for example, innovation cannibalizes

monopoly profit (e.g., Arrow (1962)), preemptive investments deter entry (e.g., Gilbert and

Newbery (1982)), or firms innovate to escape competitive pressure (e.g., Aghion et al (2005)).

The institutional details of the market matter precisely because there is no single theory.

The more recent empirical literature accordingly explores more well-defined settings (e.g.,

Cockburn and Henderson (1994); Nickell (1996); Lerner (1997); Blundell et al (1999); Aghion

et al (2004); Aghion et al (2009); Goettler and Gordon (2011)).

We examine a wave of new technology adoption in the portland cement industry that

plays out over roughly forty years. Portland cement is a finely ground dust that forms

concrete when mixed with water and coarse aggregates such as sand and stone. Production

involves feeding limestone and other raw materials into large rotary kilns that reach peak

temperatures of 1400-1450◦ Celsius. Our focus is on the adoption of precalciner technology.

Precalciners attach to one end of the rotary kiln and preheat raw materials using exhaust

gases from the kiln and a supplementary combustion chamber, and subsequently reduce fuel

costs by 25 to 35 percent relative to the dominant incumbent kiln technologies. Installation

is not simply a matter of appending to an existing rotary kiln, however, as precalciners alter

the requisite kiln properties by reducing the time that raw materials must spend in the kiln.

Thus, the adoption of precalciner technology requires the construction of an entirely new

kiln system, and this entails substantial capital outlays (e.g., Ryan (2012)).

To inform the empirical inquiry, we first adapt the Dasgupta and Stiglitz (1980) model

of Schumpeterian competition. With our modification, a focal firm can invest in cost-

reducing technology by paying a pre-determined investment cost. It then competes a la

Cournot against its competitors. Investment incentives increase with the cost reduction

1Aghion and Tirole (1994) refer to the Schumpeterian hypotheses regarding the impact of firm size andmarket structure on innovation as the second most tested relationships in industrial organization. Even thenumber of literature reviews is daunting (see e.g., Kamien and Schwartz (1982); Baldwin and Scott (1987);Cohen and Levin (1989); Cohen (1995); Gilbert (2006); Cohen (2010)).

1

available and, as in Dasgupta and Stiglitz (1980), decrease with the number of competitors.

The model clarifies, however, that the number of competitors affects investment decisions

only through its interaction with the cost reduction. The intuition is straight-forward: if the

benefits of technology adoption are small then there is little remaining scope for deterrents

such as competition to have an incremental impact. By contrast, if the benefits of adoption

are meaningful then the degree of competition helps determine whether an adequate return

on investment can be earned. The immediate implication is that empirical estimates of the

impact of competition on innovation are likely to suffer from misspecification bias unless

the regression accounts for this interaction. This can be done either by (i) using settings in

which benefits are invariant, or (ii) exploiting independent variation in competition and the

benefits of innovation.2

We take the second approach. The portland cement industry provides an advantageous

laboratory for two reasons. First, we are able to observe new technology adoption directly,

as well as the economic circumstances of cement firms. This sidesteps the use of patent

counts and R&D expenditures as proxies for innovation, a common practice in the empirical

literature (Gilbert (2006); Cohen (2010)). Second, the fuel cost advantage of precalciner

technology is measurable. We calculate the “fuel cost gap” as the difference between the

fuel costs of incumbent kilns and those of the technological frontier. Fluctuations in fossil

fuel prices create variation in the fuel cost gap over the sample period of 1974-2010. Be-

cause transportation costs create localized cement markets (Miller and Osborne (2014)),

the fluctuations are experienced by firms in quite different competitive positions. These

complementary sources of variation allow us to capture empirically the interactive effects of

competition. To our knowledge, we are the first in the empirical literature to examine explic-

itly the interaction of competition with the magnitude of the new technology opportunity.

We use reduced-form logit regressions to model the adoption decision. The regression

sample includes 7,450 kiln-year observations, including data from 445 unique kilns that could

be replaced with precalciner technology. The baseline specification includes as regressors the

fuel cost gap and its interactions with two variables that measure the amount of domestic

and import competition, respectively. The results are entirely consistent with the predictions

of the theoretical model: the probability of precalciner adoption increases with the fuel cost

gap, and this relationship dissipates as domestic and import competition intensify. Overall,

2Substantial empirical literatures examine how opportunity and appropriability affect innovation andnew technology adoption (Cohen (2010)). The analogs to these concepts in our model are the marginalcost reduction and the number of competitors, respectively. Generalizing somewhat, our theoretical resultsidentify an interaction between opportunity and appropriability that is relevant for empirical research.

2

we estimate that the median effect of a one standard deviation increase in the fuel cost gap

raises the probability of precalciner adoption by 28%. However, among kilns in the lowest

quartile of our domestic competition variable the median effect is 41%, while among kilns

in the highest quartile the median effect is 15%. The results are statistically significant and

robust across a range of specification and modeling choices.

Further, when we impose that domestic and import competition affect precalciner

adoption directly, rather than through their interactions with the fuel cost gap, statistical

significance diminishes greatly and is sensitive to specification choices. This underscores the

importance of accounting for the “confounding factors” that Gilbert (2006) and Cohen (2010)

characterize as plaguing the early empirical literature. It also reinforces the connection be-

tween the adapted Dasgupta and Stiglitz (1980) model and the empirical setting. Interpreted

through the lens of the theoretical model, our results indicate that competition limits the

ability of firms to appropriate the benefits of new technology adoption. This appears to dom-

inate impacts documented in other market settings that arise due to strategically-motivated

investments (e.g., Grabowski and Baxter (1973); Cockburn and Henderson (1994); Lerner

(1997)) and investments to “escape the competition” (e.g., Aghion et al (2004, 2009)). Inter-

estingly, our results are closer to the finding of Goettler and Gordon (2011) that the presence

of AMD as a competitor in the computer microprocessor industry diminishes Intel’s R&D

spend, despite the stark differences between the two empirical settings.

Our results have practical implications for policy-makers. The production of portland

cement accounts for around 5% of global anthropogenic CO2 emissions (Van Oss (2005)). Be-

cause demand is inelastic at prevailing prices (Miller and Osborne (2014)), whether market-

based regulation would produce abatement from the industry depends, in large part, on

whether it would spur plants to adopt less carbon-intensive production technologies. The

existing empirical studies on the market-based regulation of the cement industry do not ad-

dress this margin (e.g., Fowlie et al (2015); Miller et al (2015b)).3 Our results indicate that

competitive considerations are likely to slow the pace of technology adoption, and raise the

possibility that direct subsidies could amplify the efficacy of market-based regulation.4

The rest of this paper is organized as follows. Section 2 highlights the empirical context,

by describing the changes in kiln technology in the portland cement industry over the sample

3The analogous question has been researched in the context of the automobile sector (e.g., Knittel (2011);Linn and Klier (2012)).

4The practical implications extend to a number of other public policy interventions. For instance, theantitrust agencies often cite innovation as a consideration in their merger challenges (Gilbert (2006)). Ourresults indicate that such judgments are appropriate only on a case-by-case basis, as increases in marketpower can in some instance result in more innovation and technological adoption.

3

period. Section 3 presents the adapted Dasgupta and Stiglitz (1980) model of Schumpeterian

competition, and discusses the implications for empirical research. Section 4 describes the

data sources and defines the dependent, independent and control variables. Section 5 presents

the empirical model and discusses identification. Section 6 presents the econometric results,

and Section 7 concludes.

2 The Portland Cement Industry

We examine the portland cement industry in the United States over the period 1974-2010.

Table 1 tracks the progress of technology adoption over the sample period. At the outset,

nearly all kilns are classified as relatively inefficient wet kilns or long dry kilns.5 Only a

single precalciner kiln is operational in 1974, and a handful of “preheater” kilns account for

the remainder of industry capacity.6 Yet precalciner kilns become the dominant production

technology by the end of the sample period. In 2010, the last year of data, precalciner kilns

account for fully 70% of industry capacity. By contrast, fewer than 60 wet and long dry kilns

remain operational, and those older kilns account for only 17% of industry capacity. As new

technology replaces old, total capacity increases even as the number of kilns falls because

precalciner kilns have greater capacities.

The wave of precalciner adoption coincides with variation in underlying economic con-

ditions. Time-series fluctuations in different fossil fuel prices creates variation in the fuel

cost advantage of precalciners. Figure 2 shows the average fuel costs for each of the kiln

technologies, using snapshots across the sample period. We calculate fuel costs based on the

kiln efficiency and the price of the kiln’s primary fossil fuel (see Appendix A for details).

Within each technology class, average fuel costs decrease and then increase through the

sample period. The figure also provides the fuel costs of the technological frontier, which we

define as the combination of precalciner efficiency and the cheapest fossil fuel. The “fuel cost

gap” between older kilns and the frontier plays an important role in the empirical analysis.

The fuel cost gap tends to be larger when fossil fuel prices (and thus fuel costs) are high.7

The final column of the table provides the national average price of portland cement, in

5The distinction between “wet” and “dry” processes arises from how raw materials are processed: withwet kilns, the raw materials are wet-ground to form a slurry; with dry kilns, raw materials are dry-groundto form a powder. Modern precalciner kilns use the dry process.

6In preheater kilns, the raw materials are preheated using only exhaust gases; a supplementary combustionchamber is not employed.

7There is a well known analogy in the automobile industry: the driving cost of vehicles with low miles-per-gallon (MPG) is more sensitive to the gasoline price than that of high MPG vehicles, and automobileprices adjust accordingly (e.g., Busse, Knittel and Zettelmeyer (2013); Langer and Miller (2013)).

4

Table 1: Plants, Kiln Technologies, and Capacity Shares

Wet Long Dry Dry with Dry with Total TotalYear Kilns Kilns Preheater Precalciner Kilns Capacity

1974 230 58% 157 33% 25 8% 1 1% 413 79.281975 218 56% 145 32% 31 11% 1 1% 395 79.481980 167 49% 102 27% 38 16% 10 8% 317 76.471985 107 36% 81 24% 36 16% 24 24% 248 76.511990 79 32% 67 23% 38 19% 24 27% 208 73.071995 71 28% 67 22% 38 19% 27 30% 203 75.662000 65 24% 62 20% 35 18% 35 39% 197 83.622005 46 15% 48 14% 38 17% 49 54% 181 93.972010 28 8% 29 9% 32 14% 64 70% 153 103.48Notes: The table shows kiln counts, capacity shares, and total capacity in five-year snapshots spanning 1974-2010.Tabulations are provided separately for each of the four production technologies: wet kiln, long dry kilns, dry kilnswith preheaters, and dry kilns with precalciners. Total capacity is in millions of metric tonnes. The data are for thecontiguous U.S. and are obtained from the PCA Plant Information Survey.

Table 2: Fuel Costs per Metric Tonne of Cement

Wet Long Dry Preheater Precalciner Frontier Price

1974 26.66 23.81 19.38 20.33 12.02 98.981975 31.23 28.25 21.70 22.41 15.36 106.491980 32.42 25.76 23.21 18.67 18.67 122.161985 23.08 18.65 14.57 13.53 13.53 95.871990 15.73 13.13 10.90 10.23 9.59 86.871995 12.75 10.11 8.61 8.00 6.06 95.432000 11.13 9.59 7.49 7.27 6.75 101.742005 14.32 12.28 9.88 9.50 7.44 104.262010 17.91 15.67 12.19 11.72 11.58 92.00

Notes: The table provide average fuel costs across four different kiln technologies (wet, longdry, preheater, and precalciner), along with the average price of portland cement. Fuel costsare based on authors’ calculations. See Appendix A for details. The “frontier” is calculatedbased on the fuel efficiency of precalciner kilns and the most affordable fossil fuel. Nationalaverage prices are obtained from the USGS Minerals Yearbook. All numbers are in real 2010dollars per metric tonne of cement output.

order to establish that fuel costs absorb a substantial portion of plant revenues, especially

during periods of high fossil fuel prices.

These fuel cost changes are experienced by plants in an array of different competition

conditions. Cement plants sell to ready-mix concrete producers and large construction firms.

5

Figure 1: Cement Plants in 2010

Most cement is trucked directly from the plant to the customer, though some cement is

first transported by barge or rail to distribution terminals and then trucked to customers.

The costs of transportation are large relative to the overall price, which creates localized

market power. Figure 1 provides a map of the operational cement plants circa 2010. Some

geographic areas have many plants (e.g., southern California) while others areas have only

a single nearby plant (e.g., South Dakota). For perspective, the distance between plants

and customers often is less than 100 miles, and only rarely more than 200 miles (Miller and

Osborne (2014)). Thus, plants differ substantially in the number of relevant competitors

that they face. They also differ in their proximity to the customs districts through which

foreign imports can enter the U.S. market, and our empirical analysis attempts to analyze

the effects of both domestic and import competition.

3 Theoretical Rationale

Our core methodological insight that the impact of competition on new technology adoption

depends on the magnitude of the technological opportunity. We develop this interaction in

our empirical results, exploiting variation in the fuel cost gap and local competitive condi-

tions. We also show that competition does not robustly predict precalciner adoption, except

through its interaction with the fuel cost gap. Here we develop a theoretical rationale for

the interactive effects of competition and opportunity.

6

To this end, we adapt an oligopoly model initially studied in Dasgupta and Stiglitz

(1980) that generates Schumpeterian predictions.8 The model features N firms engaged in

Cournot competition. A single focal firm can pay an investment cost to lower its marginal

costs. Both the investment cost and the marginal cost reduction are fixed – the focal firm

can pay the investment cost, or not. This represents a key departure from the original

model, in which investment is treated as a continuous variable.9 With the modification, the

interactive effects of competition and the technological opportunity become transparent. The

modification also accords with our empirical application, in which cement plants consider

whether to install a technology that lowers production costs by a known amount.

The model has two stages: First, the focal firm chooses whether to reduce its marginal

costs by investing in the new technology. Second, payoffs are determined according to

Cournot principles. We solve the model with backward induction. We then develop com-

parative statics that show how competition affects new technology adoption.

3.1 Cournot equilibrium

In the second stage of the model, i = 1, 2, . . . , N firms compete in a Cournot output game.

Marginal costs are given by the vector c = (c1, c2, . . . , cN). The unweighted average of the

firms’ marginal costs is given by c = 1N

∑Ni=1 ci. Each firm i selects output qi to maximize

profit, taking as given the output of other firms. Price is determined by an inverse demand

curve P (Q) = a − Q, where Q =∑

i qi. Profit is given by πi = (P (Q) − ci)qi. The model

is standard and solutions can be derived with the usual steps. Under normalcy conditions

that guarantee positive output for each firm, equilibrium is characterized by the quantities

q∗i =(a− ci +N(c− ci))

(N + 1), (1)

8See also chapter five of Stiglitz and Greenwald (2014). The model is featured in the Gilbert (2006)discussion of innovation incentives with non-exclusive property rights.

9In the original Dasgupta and Stiglitz (1980) model, each firm can invest to lower marginal costs. Theamount invested determines the magnitude of marginal cost reductions. The equilibrium level of invest-ment diminishes as the number of firms increases, which derives from two intuitions: First, the benefit ofcost-reducing technology is proportional to the amount of production such that, for example, a monopolistbenefits more than a (single) duopolist. This follows the basic insight of Arrow (1962). Second, competitionhampers the ability of firms to appropriate the benefits of cost-reducing technology, amplifying the discrep-ancy between the social and private returns to investment. The end result are predictions that accord withSchumpeterian hypotheses regarding competition and innovation.

7

and this corresponds to the following market prices and total output:

P ∗ =1

N + 1a+

N

N + 1c Q∗ =

N

N + 1(a− c). (2)

A number of initial comparative statics are useful in solving the first stage of the model.

We start with three equilibrium pass-through relationships:

∂P ∗

∂ci=

1

N + 1

∂Q∗

∂ci=−1

N + 1

∂q∗i∂ci

= − N

N + 1(3)

Thus, the pass-through a firm-specific cost change to the market price is determined by the

number of firms, equaling 1/2 in the special case of monopoly and converging to zero as N

grows large. Next, with an application of the envelope theorem, tractable expressions can

be obtained that relate marginal costs and the number of firms to profit:

∂πi∂ci

= − 2N

N + 1q∗i

∂πi∂N

= − 2a

(N + 1)2q∗i

∂2πi∂ci∂N

=2Na

(N + 1)3(4)

The profit of each firm decreases with its marginal costs and N . The negative effect of a

firm’s marginal costs on its profit attenuates as N grows large, consistent with the benefit

of cost-reducing technology being proportional to the amount of production. Analogously,

the negative effect of competition also attenuates as ci grows large because high-cost firms

have small levels of output and profit.

3.2 Technology adoption

We now turn to the first stage of the model. Suppose that firm i can purchase a cost-

reducing technology by paying the investment cost I. The technology, if adopted, lowers

the marginal costs of firm i by ∆c. Let the marginal costs with and without adoption be c1i

and c0i , respectively, such that c1

i = c0i − ∆c. The corresponding levels of profit (excluding

the investment cost) are πi(c1i , c−i) and πi(c

0i , c−i), where c−i is a vector of competitor costs.

Technology adoption occurs if and only if I < πi(c1i , c−i)− πi(c0

i , c−i).

We use the following equality to develop the comparative statics of interest to the

empirical application:

πi(c1i , c−i)− πi(c0

i , c−i) = − πi(ci, c−i)

∂ci

∣∣∣∣ci=c0i

∆c (5)

8

The equality holds because equilibrium profit is a quadratic function of marginal costs, given

Cournot competition, linear demand, and constant marginal costs.10 The expression greatly

simplifies the analysis, and allows the implications of the model for technology adoption to

flow easily from the Cournot comparative statics in equations (3) and (4).

Result 1: The incentive to adopt the cost-reducing technology increases with ∆c, and the

magnitude of this effect decreases with N .

Proof: Differentiating the RHS of equation (5) with respect to ∆c yields −∂πi/∂ci, which is

positive by equation (4). Next, differentiating the RHS of equation (5) with respect to ∆c

and N yields −∂2πi/(∂ci∂N) which equals −2Na/(N + 1)3 by equation (4).

Result 2: The incentive to adopt the cost-reducing technology decreases with N , and the

magnitude of this effect is proportional to ∆c.

Proof: Differentiating the RHS of equation (5) with respect to N yields − ∂2πi∂ci∂N

∆c, which

equals − 2Na(N+1)3

∆c by equation (4). This is negative and proportional to ∆c.

These results indicate an interactive effect of the technological opportunity (∆c) and

the degree of competition (N). The intuition is straight-forward: if the benefits of new

technology adoption are small then there is little remaining scope for deterrents such as

competition to have incremental impact. This is why competition affects adoption incentives

only through its interaction with the technological opportunity. By contrast, if the benefits

of adoption are large then competition affects whether an adequate return is possible.

3.3 Discussion

The theoretical analysis provides direction for empirical research. To identify the effects of

competition on new technology adoption, a suitable experiment should seek to either hold

the technological opportunity fixed, or to incorporate the opportunity into the analysis by

modeling its interactions. The difficulties associated with implementing such an experiment

are widely recognized in the academic literature (e.g., Gilbert (2006)). Most fundamentally, it

is exceedingly rare to find empirical settings that feature variation in competitive conditions

with technological opportunity that is either fixed or measurable.11

10The right-hand-side of equation (5) represents a single step of Newton’s method. This is sufficient for afirst order approximation under general conditions, and is exact in the quadratic case. See Jaffe and Weyl(2013) and Miller, Remer, Ryan and Sheu (2015a) for other economic applications of this technique.

11This may help explain why the positive relationship between market concentration and R&D spendingidentified in the early empirical literature, based on cross-industry comparisons, tend to disappear once

9

Our empirical application, to which we return next, overcomes these difficulties using

the institutional details of the portland cement industry outlined in Section 2. In particular,

we exploit that fossil fuel price fluctuations create measurable variation in the fuel cost gap

between older kilns and technological frontier, and that these changing technological oppor-

tunities are experienced by firms in quite different competitive situations. This identifies

the interactive effects of competition in a manner that, to our knowledge, is unique in the

substantial empirical literature on competition and new technology adoption.

4 Data and Variables

4.1 Data sources

We draw on data from numerous sources to construct a panel of kiln-year observations

that span the contiguous United States over 1974-2010. The kiln data are from the Plant

Information Survey (PIS), an annual publication of the Portland Cement Association (PCA).

This publication provides the location, owner, and primary fuel of each cement plant in the

U.S. and Canada, as well as the age, annual capacity and technology of each kiln. The

PIS is published annually over 1974-2003 and semi-annually in 2004, 2006, 2008 and 2010.

We impute values in the missing years based on the data from the preceding and following

years.12 There are 4,416 plant-year observations and 8,775 kiln-year observations in the

sample.

We calculate the fuel costs of production based on kiln efficiency and fossil fuel prices.

The details of the calculation are deferred to Appendix A. We use the PCA’s U.S. and Cana-

dian Portland Cement Labor-Energy Input Survey to measure kiln efficiency. The survey is

published intermittently, and we use 1974-1979, 1990, 2000, and 2010. We use three sources

to obtain fossil fuel prices. We obtain the national average delivered bituminous coal price

in the industrial sector over 1985-2010 from the annual Coal Reports of the Energy Informa-

industry-level control variables are added to the specification (e.g., Scott (1984); Levin, Cohen and Mowery(1985)). ? (?) conclude that “[t]he most common feature of the few R&D and innovation analyses thathave sought to control for the underlying technological environment is a dramatic reduction in the observedimpact of the Schumpeterian size and market power variables.”

12In most instances, imputation is trivial because capacity, process technology and fuel are persistent acrossyears. When the data from the preceding and following years differ, we use the data from the precedingyear. We are able to identify kilns that are built in the missing years because the PIS provides the year ofconstruction for each kiln. A handful of kilns in the PIS drop out and then reappear in latter years. Wetreat these observations on a case-by-case basis leveraging detailed qualitative and quantitative informationprovided in the Minerals Yearbook of the United States Geological Survey (USGS), an annual publicationthat summarizes a census of portland cement plants.

10

tion Agency (EIA). We backcast these prices to the period 1974-1984 using historical data

on national average free-on-board prices of bituminous coal published in the 2008 Annual

Energy Review of the EIA.13 We obtain the national prices of petroleum coke, natural gas,

and distillate fuel oil – again for the industrial sector – from the State Energy Database

System (SEDS) of the EIA. The SEDS also includes data on coal prices, but no distinction

is made between bituminous coal, sub-bituminous coal, lignite, and anthracite, despite the

wide price differences that arise between those fuels.

The demand for portland cement derives from construction activity. We use county-

level data from the Census Bureau on construction employment and building permits to

account for demand-side fluctuations.14 Construction employment is part of the County

Business Patterns data. We identify construction as NAICS Code 23 and (for earlier years)

as SIC Code 15. The data for 1986-2010 are available online.15 The data for 1980-1985 are

obtained from the University of Michigan Data Warehouse. The building permits data are

maintained online by the U.S. Department of Housing and Urban Development.16

Lastly, we use the Minerals Yearbook, an annual publication of the USGS, to account

for the presence of foreign imports. The availability of imported cement itself is due primar-

ily to transoceanic freighter technology adopted around 1980; in subsequent years imports

are common especially when demand outstrips domestic capacity. The Minerals Yearbook

identifies the customs districts through which foreign importers enter the domestic mar-

ket. The number of customs districts grows during the sample period. We focus on active

customs districts, defining each district as active in the first year in which imports exceed

five thousand metric tonnes, and continuing through all subsequent years.17 The Minerals

Yearbook also is the source of the national price data provided in Table 2.

13The coal price data are in dollars per metric tonne, and we use the conversion factor of 23 mBtu permetric tonne to convert the data to dollars per mBtu. This conversion factor reflects the average energycontent of bituminous coal obtained by cement plants based on the labor-energy input surveys.

14For both the construction employment and building permits, it is necessary to impute a small numberof missing values. We calculate the average percentage difference between the observed data of each countyand the corresponding state data, and use that together with the state data to fill in the missing values.

15See http://www.census.gov/econ/cbp/download/, last accessed April 16, 2014.16See http://socds.huduser.org/permits/, last accessed April 16, 2014.17The two exceptions are Duluth, Minnesota and Milwaukee, Wisconsin, which we code as inactive starting

in 2005 due to the complete cessation of importing.

11

0

.02

.04

.06

.08

Den

sity

of P

reca

lcin

er A

dopt

ion

1975 1980 1985 1990 1995 2000 2005 2010

Figure 2: Empirical Distribution of Precalciner AdoptionNotes: The histogram is based on 7,540 kiln-year observations over 1974-2008. There are 445 unique wet,long dry, and preheater kilns in the sample, of which 132 are replaced with precalciner technology.

4.2 Variable definitions

4.2.1 Dependent variable

Our dependent variable captures the adoption of precalciner technology. We create an indi-

cator variable at the kiln-year level that equals one under the joint conditions that (i) it is the

final year of production for the kiln, and (ii) the plant constructs a precalciner kiln within a

four-year window starting the prior year. These timing assumptions help capture instances

in which the replacement of older kilns is not immediate. However, most replacements are

immediate and the empirical results are virtually unchanged if alternative windows are used.

Our approach restricts the baseline regression sample to wet, long dry, and preheater kilns

in the years 1974-2008. There are 7,540 such kiln-year observations, featuring 445 unique

kilns that could be upgraded to precaliner technology.

Figure 2 plots the empirical distribution of the dependent variable over time. Overall,

we observe 132 instances in which an older kiln is replaced with precalciner technology.

The number of upgrades exceeds the total number of precalciner kilns (e.g., see Table 1)

because plants sometimes replace multiple wet or long dry kilns with a single precalciner

kiln. Adoption predominately occurs during the first decade of the sample (i.e., 1974-1983)

and during the last decade of the sample (i.e., 1999-2008). These periods coincide with

higher fossil fuel prices. Thus, the timing of precalciner adoptions is consistent perhaps the

most basic prediction of the theoretical model: the incentive for new technology adoption is

amplified when the new technology provides greater benefit.

12

4.2.2 Competition and fuel costs

We measure the competition that a kiln faces using two variables that reflect the degree of

domestic and import competition, respectively. The first, which we refer to as Rival Capacity,

is the total capacity at competing plants within 400 miles. Because the variable increases

with the both number and capacity of competitors, it builds on the theoretical model by

incorporating that larger plants are likely to exert more competitive pressure than smaller

plants. That said, our empirical results are robust to an alternative competition measure

based on competitor counts. We select the baseline distance threshold of 400 miles based

on the finding that cement rarely is shipped farther than 200 miles (Miller and Osborne

(2014)), which implies that plants separated by more the 400 miles are unlikely to compete

for customers. We also develop results using alternative distance thresholds.

We measure import competition using the variable Import Proximity, defined as the

inverse distance between the kiln and the nearest active customs district (at which foreign

imports can arrive). The variable equals one for a kiln located at the customs district, and

decreases as the miles between the kiln and the customs district grows large. Because imports

become economically viable in 1980 due to innovations in transoceanic freighter technology,

we set Import Proximity to zero over the period 1974-1979. The relevant distance changes

over the sample period, even for a given kiln, as the set of active customs districts expands.

The theory indicates that competition affect the new technology adoption through its

interaction with the magnitude of the technological opportunity. To measure this oppor-

tunity, we first calculate the fuel costs of the technological frontier based on precalciner

technology and the cheapest fossil fuel. We then define the variable Fuel Cost Gap as the

difference between a kiln’s fuel costs and those of the technological frontier. Appendix A

contains details how we calculate fuel costs. The three main regressors of interest are Fuel

Cost Gap, which should positively affect precalciner adoption, and the interactions Fuel Cost

Gap×Rival Capacity and Fuel Cost Gap×Import Proximity, which should be deterrents.

4.2.3 Control variables

We use a number of control variables in our regressions. These include kiln age, kiln capacity,

fixed effects for the kiln technology (i.e., indicators for wet, long dry and preheater kilns),

and a linear time trend. The definitions of these variables are straight-forward. We also

define a variable, Demand, that captures the demand for portland cement among nearby

counties. The variable is a linear combination of construction employment and building

permits, which we observe at the county-year level. The combination is based on a regression

13

of state-level consumption, which is available in the Minerals Yearbook, on construction

employment and building permits aggregated to the state-level. These two regressors have a

great deal of explanatory power in the consumption regression: the resulting R2 is 0.89. We

apply the regression coefficients to the county-level data on construction employment and

building permits to obtain county-level predictions on consumption. The specific formula is

CONS = 0.018×PER+0.012×EMP where CONS is county-level predicted consumption,

PER is building permits, and EMP is construction employment. The control variable

Demand then is defined for each kiln-year observation as the sum of predicted consumption

among counties within 400 miles. The main results hold when construction employment and

building permits are included independently, but the high degree of correlation between the

two control variables makes interpretation of their coefficients difficult.

5 Empirical Model

5.1 Framework

Our estimation strategy follows the two-stage approach of the theoretical model. In the first

stage, producers determine whether to adopt precalciner technology. In the second stage,

these producers compete, taking the outcomes of the first stage as given. We conceptualize

producers as playing this two-stage game each year (or period), so as to exploit the annual

observations in our panel data. We assume that producers do not consider how their decisions

affect subsequent periods. Thus, the framing is analogous to the static perfect information

games of entry estimated elsewhere in the literature (e.g., Bresnahan and Reiss (1991); Berry

(1992); Toivanen and Waterson (2005); Ciliberto and Tamer (2009); Perez-Saiz (2015)). This

approach avoids the simplifying restrictions that would be required to accommodate forward-

looking producers.18

Denote the profit generated by kiln j in the second stage of period t as Πijt, where the

superscript i equals 1 if precalciner technology is adopted in the first stage, and 0 otherwise.

We decompose this profit as follows:

Πijt = πi(x′jtβ) + εijt, (6)

18Consider that the “state space” of the industry is high-dimensional and includes at least the locationand owner of each kiln, the kiln technologies, local demand conditions, import proximities, and fossil fuelprices. Our approach allows us to capture empirically the most relevant salient market forces. Dynamicestimation would require the sacrifice of this realism to reduce the dimensionality of the state space.

14

The functions πi() determine how profit is affected by fuel costs and competition: xjt is a

vector of regressors and β is a vector of parameters. The fixed costs associated with the

decision i ∈ {0, 1} are captured by the profit shock εijt, which we assume is unobservable

and additively separable from the variable profit function. To estimate the reduced-form,

we impose the normalization that π0jt = 0 and assume that fixed costs are stochastic with

an extreme value distribution. The probability of adoption is

Pr(Adoption) =exp(π1(x′jtβ)

1 + exp(π1(x′jtβ)(7)

We estimate the logit model with standard maximum likelihood techniques, and cluster

standard errors at the kiln level to account for autocorrelation and heteroskedasticity.

The error term in the regression has the interpretation of being an unobserved driver

of precalciner adoption. The key identifying assumption, then, is that this error term is

orthogonal to fuel costs and our measures of competition. The exogeneity of fuel costs

seems straight-forward. Fossil fuel prices do not follow the pro-cyclical pattern of cement

consumption (e.g., see Figure B.2), cement accounts for a small fraction of the fossil fuels

used in the U.S., and there is no reason to suspect that fuel prices would be correlated with

labor expenses or other costs of production. The exogeneity of domestic competition is more

difficult to establish, and usual the cost and demand instruments are invalid when estimating

profit functions. However, if some regions have unobservable and persistent profit shocks

(e.g., high demand or low costs) then this would bias the regression against the result that

domestic competition is a deterrent to precalciner adoption.19

5.2 Summary Statistics

Table 3 provides summary statistics and a matrix of correlation coefficients. We highlight

that the mean value of the dependent variable, labeled Adoption, is 0.018. Thus, despite the

fact that 132 of the 445 unique kilns in the regression sample are upgraded to precalciner

technology, the chance of upgrade in a specific year is less than two percent for the average

19There are a number of other reasons to think that our results are not subject to strong omitted variablebias. First, the results indicate that the impacts of domestic competition and foreign competition (which ismore clearly exogenous) are qualitatively similar. Second, we obtain nearly identical results when using atwo-year lag of Rival Capacity as an instrument. While this procedure would only partially mitigate biasdue to persistent unobserved profit shocks, the fact that results are unchanged indicates that perhaps bias isnot present. Third, given that the baseline specification includes a control for demand that explains nearly90% of state-level consumption, it is plausible that the error term is dominated by idiosyncratic variation infixed costs and investment costs that is unrelated to the competitive structure of the industry.

15

kiln-year observation. This has implications for how to interpret the marginal effects of the

regressors, as even small effects become substantial when aggregated over time.

The average kiln is 30 years old with an annual capacity of 250 thousand metric tonnes.

Its fuel cost gap is $6.74, and it is located such that there are nearly 12MM metric tonnes of

rival capacity and 13MM metric tonnes of predicted consumption within 400 miles. The Fuel

Cost Gap variable and its interaction with Rival Capacity have a relatively high correlation

coefficient of 0.821. To investigate the possibility that multicollinearity could distort stan-

dard errors in estimation, we calculate the variance inflation factor (VIF) of each regressor.20

An econometric rule of thumb is that multicollinearity is potentially problematic with VIF

statistics that exceed ten (e.g., Kutner, Nachtsheim and Neter (2004)). Our regressors do

not reach this threshold.

20The VIF is calculated by (i) regressing the variable of interest on the other explanatory variables; (ii)obtaining the R2 from that regression; and (iii) applying the formula V IF = 1/(1−R2). The VIF informshow much smaller the standard error on the coefficient of the regressor would be if that regressor wereuncorrelated with other explanatory variables.

16

Tab

le3:

Sum

mar

ySta

tist

ics

and

Cor

rela

tion

Mat

rix

Corr

elati

on

Coeffi

cien

tsM

ean

St.

Dev

.V

IF(1

)(2

)(3

)(4

)(5

)(6

)(7

)(8

)(9

)

(1)

Ad

opti

on0.

018

(0.1

31)

·1.0

00

(2)

Fu

elC

ost

Gap

6.74

(6.3

4)4.5

20.0

88

1.0

00

(3)

Fu

elC

osts

Gap×

79.3

1(9

2.44

)3.4

10.0

47

0.8

21

1.0

00

Riv

alC

apac

ity

(4)

Fu

elC

osts

Gap×

0.17

(1.5

3)1.1

0-0

.003

0.2

23

0.0

98

1.0

00

Imp

ort

Pro

xim

ity

(5)

Dem

and

12.6

6(7

.28)

1.3

00.0

10

-0.2

35

-0.0

71

0.0

11

1.0

00

(6)

Kil

nA

ge29

.87

(15.

85)

1.8

30.0

81

-0.0

62

-0.0

40

0.0

84

0.1

43

1.0

00

(7)

Kil

nC

apac

ity

0.26

(0.2

6)1.6

5-0

.046

-0.2

27

-0.1

62

-0.0

16

0.1

12

-0.4

28

1.0

00

(8)

Riv

alC

apac

ity

11.8

9(5

.65)

·-0

.011

-0.0

23

0.4

27

-0.0

86

0.2

98

0.0

54

0.0

69

1.0

00

(9)

Imp

ort

Pro

xim

ity

0.01

7(0

.082

)·

-0.0

02

0.1

07

-0.0

09

0.7

99

0.0

11

0.0

84

-0.0

16

-0.1

59

1.0

00

Note

s:B

ase

don

the

regre

ssio

nsa

mp

leof

7,5

40

kiln

-yea

rob

serv

ati

on

sover

1974-2

008.Adoption

isan

ind

icato

rth

at

equ

als

on

ed

uri

ng

the

fin

al

yea

rof

akil

n’s

op

erati

on

ifth

ekiln

isre

pla

ced

wit

hp

reca

lcin

erte

chn

olo

gy

wit

hin

afo

ur-

yea

rw

ind

ow

.Fuel

Cost

Gap

isth

ed

iffer

ence

bet

wee

nth

ekil

n’s

fuel

cost

san

dth

at

of

the

tech

nolo

gic

al

fronti

er.RivalCapa

city

isth

eto

tal

com

pet

itor

cap

aci

tylo

cate

dw

ith

in400

mil

esof

the

kiln

.Im

port

Proximity

isth

ein

ver

sed

ista

nce

bet

wee

nth

ekiln

an

dth

en

eare

stact

ive

cust

om

sd

istr

ict.

Dem

and

isto

tal

pre

dic

ted

con

sum

pti

on

am

on

gco

unti

esw

ith

in400

miles

of

the

kiln

.Fuel

Cost

Gap

isin

real

2010

dollars

per

met

ric

ton

ne,

wh

ileRivalCapa

city

,Dem

and,

an

dKiln

Capa

city

are

inm

illion

sof

met

ric

ton

nes

.V

IFst

ati

stic

sare

rep

ort

edfo

rea

chvari

ab

lein

the

base

lin

esp

ecifi

cati

on

.

17

Table 4: Main Logit Regression Results

(i) (ii) (iii) (iv) (v)

Fuel Cost Gap 0.118*** 0.108*** 0.114*** 0.107*** 0.072***(0.017) (0.015) (0.015) (0.016) (0.012)

Fuel Cost Gap × -0.004*** -0.003** -0.004*** -0.003***Rival Capacity (0.001) (0.001) (0.001) (0.001)

Fuel Cost Gap × -0.131*** -0.112** -0.132*** -0.107**Import Proximity (0.034) (0.036) (0.032) (0.032)

Control Variables yes no yes yes yesTime Trend yes yes no yes yes

Psuedo-R2 0.079 0.047 0.071 0.075 0.072

Notes: The regression sample includes 7,540 kiln-year observations over 1974-2008. The dependentvariable is an indicator that equals one if during the year that the kiln is replaced for precalcinertechnology. The technology fixed effects include indicators for wet and long dry kilns. Controlvariables include Kiln Age, Kiln Capacity, Demand, and indicators for wet and long dry kilntechnology. A linear time trend also is included where indicated. The standard errors, shownin parentheses, are clustered at the kiln level to account for autocorrelation and heteroskedastity.Statistical significance at the 10%, 5%, and 1% levels are denoted with *, **, and ***, respectively.

6 Results

6.1 Main regression analysis

Table 4 provides summarizes the main regression results. Column (i) shows the baseline

specification. The main coefficients are statistically significant and take the signs predicted

by the theoretical model. The Fuel Cost Gap coefficient indicates that the adoption of

precalciner technology is positively related to the fuel cost benefits of adoption. The negative

coefficients on the interaction terms show that this relationship is mitigated both by domestic

and import competition; they also capture that the (negative) impact of competition on

adoption is larger in magnitude when Fuel Cost Gap is large. These results are robust to

different specification choices: columns (ii) and (iii) shows that the results hold when the

control variables or the time trend are excluded from the regression model, while columns

(iv) and (v) shows that the effects of domestic and import competition are independent.21

The economic implications of the coefficients are substantial. We calculate kiln-specific

effects using the baseline specification. A one standard deviation increase in Fuel Cost

21Appendix Table B.1 shows the coefficients and standard errors for the control variables. We omit themfrom the main regression table for brevity. Among the controls, only Kiln Age has a robust statisticallysignificant effect. Demand enters positively if the time trend is excluded from the specification.

18

0

.01

.02

.03

.04

Pro

babi

lity

Cha

nge

0 5 10 15 20 25Rival Capacity within 400 Miles

Figure 3: Effect of Fuel Cost Gap on the Likelihood of Precalciner AdoptionNotes: The figure shows the impact of a one standard deviation increase in Fuel Cost Gap on the likelihoodof precalciner adoption. Effects are calculated for each kiln-year observation in the regression sample andshown as blue + symbols. The solid red line summarizes how the derivatives change with Rival Capacityand is calculated using a kernel-weighted local polynomial regression.

Gap increases the likelihood of precalciner adoption by an median amount of 0.005 which,

evaluated against frequency of adoption, represents a 28% increase.22 This effect interacts

with competition. Figure 3 provides a scatterplot of the kiln-specific effects against Rival

Capacity, our measure of domestic competition. The dark red line shows the fit from a

nonparametric regression, and illustrates that the positive effect of the fuel cost gap matters

much more kilns that face little competition. Among kilns in the lowest quartile of Rival

Capacity, the median effect of a one standard deviation increase in Fuel Cost Gap is 0.007

(41%) while among kilns in the highest quartile the median effect is 0.002 (15%).

To further explore the impact of competition, we calculate how a one standard devi-

ation increase in Rival Capacity changes the likelihood of precalciner adoption. The mean

impact is a reduction in the adoption probability of 0.0035 which, evaluated against the

observed frequency of adoption, represents an 18% decrease. A tremendous amount of het-

erogeneity is present, however. Figure 4 provides a scatterplot of effect against Fuel Cost

Gap. Competition affects precalciner adoption precisely when the fuel cost benefits from

adoption are large. This heterogeneity is consistent with the theoretical model and has a

simple intuition: if the benefit of adoption is small then so is the likelihood of adoption, and

22The calculation is (0.018 + 0.0047)/0.018 = 0.28

19

0

-.05

-.1

-.15

-.2

Pro

babi

lity

Cha

nge

0 10 20 30 40 50Fuel Cost Gap

Figure 4: Effect of Rival Capacity on the Likelihood of Precalciner AdoptionNotes: The figure shows the impact of a one standard deviation increase in Rival Capacity on the likelihoodof precalciner adoption. Effects are calculated for each kiln-year observation in the regression sample andshown as blue + symbols. It is assumed that the rival capacity has fuel costs that correspond to the nationalaverage during the relevant year. The solid red line summarizes how the impact changes with Fuel Cost Gapand is calculated using a kernel-weighted local polynomial regression.

there is little remaining scope for competition to have a deterring effect.

Identifying an effect of competition on precalciner technology requires that competition

enter the specification through its interaction with the fuel cost gap. To illustrate, Table

5 shows the results of more standard “concentration-innovation” regressions, in which we

regress precalciner adoption on Rival Capacity and Import Proximity, along with Fuel Cost

Gap and the control variables. In each column, we define Rival Capacity using a different

distance thresholds. The competition coefficients are uniformly negative. Their statistical

significance is much diminished relative to the baseline specification, however, and results

are sensitive to the distance threshold. It is also possible to add Rival Capacity and Import

Proximity as regressors in the baseline specification, together with the interactions terms.

When we do so, the baseline coefficients of Table 4 are virtually unchanged, while the

statistical significance of Rival Capacity and Import Proximity again are sensitive to the

distance threshold and other specification choices.

20

Table 5: Basic Concentration Regressions

(i) (ii) (iii) (iv) (v) (vi) (vii) (viii)

Fuel Cost Gap 0.071*** 0.070*** 0.071*** 0.070*** 0.071*** 0.069*** 0.071*** 0.069***(0.012) (0.011) (0.011) (0.011) (0.011) (0.011) (0.012) (0.011)

Rival Capacity -0.020 -0.040 -0.041 -0.042* -0.035* -0.045** -0.017 -0.020(0.036) (0.036) (0.026) (0.025) (0.018) (0.019) (0.014) (0.014)

Import Proximity -1.295* -1.287 -1.424* -1.328 -1.615** -1.533* -1.505** -1.443(0.747) (0.934) (0.734) (0.904) (0.746) (0.917) (0.763) (0.936)

Psuedo-R2 0.071 0.071 0.073 0.072 0.074 0.075 0.072 0.072

Distance Threshold 200 200 300 300 400 400 500 500Diesel-Adjusted no yes no yes no yes no yes

Notes: The regression samples include 7,540 kiln-year observations over 1974-2008. The dependent variable is an indicator thatequals one if during the year that the kiln is replaced for precalciner technology. All regressions include Demand, Kiln Age, andKiln Capacity as control variables, as well as fixed effects for the kiln technology and a linear time trend. The standard errors,shown in parentheses, are clustered at the kiln level to account for autocorrelation and heteroskedastity. Statistical significance atthe 10%, 5%, and 1% levels are denoted with *, **, and ***, respectively.

6.2 Robustness analysis

Table 6 presents results from additional regressions. The specification in column (i) allows

for a nonlinear impact of domestic competition by incorporating the interactions of Fuel

Cost Gap with both Rival Capacity and its square. The signs of these coefficients are

consistent with a nonlinear effect. To evaluate further, we calculate the contribution of

domestic competition terms to the adoption incentive.23 Appendix Figure B.1 plots the

results. There is no identifiable range over which competition increases the likelihood of

adoption. Thus, the result of the baseline specification that competition mitigates adoption

incentives (with considerable heterogeneity) appears robust through the range of the data.

Column (ii) adds Rival Fuel Costs as additional regressor. The variable is defined as the

average fuel costs among competing plants within 400 miles. With some additional algebra,

the theoretical model can be shown to imply that this variable should have a positive impact

on technology adoption; the intuition is that obtaining a return on investment is easier with

less efficient competitors. The regression supports this prediction as well, which reinforces

that the main results are well motivated by the theoretical model.

Column (iii) shows results obtained with instrumental variables. Specifically, we in-

strument for Fuel Cost Gap × Rival Capacity using a two-year lags of the variable. While

23Contribution = 0.012× Fuel Cost Gap × Rival Capacity − 0.001× Fuel Cost Gap × Rival Capacity2

21

Table 6: Additional Regression Results

(i) (ii) (iii) (iv) (v) (vi) (vii)

Fuel Cost Gap 0.043 0.118*** 0.111*** 0.106*** 0.114*** 0.123*** 0.114***(0.035) (0.016) (0.020) (0.014) (0.015) (0.018) (0.015)

Fuel Cost Gap × -0.012 -0.005*** -0.005** -0.008*** -0.006*** -0.003*** -0.005***Rival Capacity (0.008) (0.001) (0.002) (0.003) (0.002) (0.001) (0.002)

Fuel Cost Gap × -0.001*Rival Capacity2 (0.000)

Fuel Cost Gap × -0.141*** -0.127*** -0.132*** -0.104* -0.107* -0.135*** -0.122**Import Proximity (0.033) (0.031) (0.038) (0.032) (0.033) (0.034) (0.034)

Rival Fuel Costs 0.097***(0.020)

Distance Threshold 400 400 400 200 300 500 400Diesel-Adjusted no no no no no no yesInstruments no no yes no no no no

Psuedo-R2 0.074 0.093 0.084 0.080 0.081 0.074 0.083

Notes: The dependent variable in all regressions is an indicator that equals one if during the year that the kiln is replaced forprecalciner technology. Column (i)-(ii) and columns (iiv)-(vii) show the results of standard logit regression, based on 7,540kiln-year observations over 1974-2008. Column (iii) shows results when Rival Capacity and Rival Fuel Costs are treated asendogenous variables: two-year lags of those variables are used as instruments. The sample in column (ii) includes 6,682kiln-year observations over 1976-2008. All regressions include Demand, Kiln Age, and Kiln Capacity as control variables,as well as fixed effects for the kiln technology and a linear time trend. The standard errors, shown in parentheses, areclustered at the kiln level to account for autocorrelation and heteroskedastity. Statistical significance at the 10%, 5%, and1% levels are denoted with *, **, and ***, respectively.

22

this procedure should only partially mitigate bias due to persistent unobserved profit shocks,

the fact that the IV results are virtually unchanged relative to the baseline indicates that

bias may not be present in the first instance. Columns (iv)-(vii) show that the main results

hold under a variety of different distance thresholds that determine the circumference at

which competing plant are incorporated into Rival Capacity.

The main results also are robust to a host of alternative modeling and specification

changes that we omit from the tables in order to conserve space. Competitors need not be

weighted by capacity in the construction of Rival Capacity, though we think such weighting

is appropriate. Additional controls can be added to the specification, including GDP growth,

the national average cement price, and state fixed effects. Fuel costs can be calculated with

an adjustment for waste fuels. The logit framework itself can be abandoned entirely in favor

of hazard rate regressions. None of these robustness checks cast doubt on the empirical

relationships obtained from the baseline regression.24

7 Conclusion

Coming soon.

24In Appendix Table B.2 we list the precalciner kilns constructed over 1974-1981. These examples ofearly adoption should occur in locales that are relatively insulated from competition, given the econometricresults. The table shows the value of Rival Capacity for each kiln alongside the average value of RivalCapacity among kiln observations from the same year. For 13 of the 17 precalciner kilns, the amount ofnearby competitor capacity indeed is lower than the corresponding national average. Thus, a simple “eyeballcheck” on the raw data corroborates the econometric analysis described above.

23

References

Aghion, Philippe and Jean Tirole, “The Management of Innovation,” Quarterly Journal

of Economics, 1994, 109, 1185–1209.

, Nicholas Bloom, Richard Blundell, Rachel Griffith, and Peter Howitt,

“Competition and Innovation: An inverted U-relationship,” Quarterly Journal of Eco-

nomics, 2005, 120, 701728.

, Richard Blundell, Rachel Griffith, Peter Howitt, and Susanne Prantl, “En-

try and Productivity Growth: Evidence from Microlevel Plant Data,” Journal of the

European Economics Association, 2004, 2, 265–276.

, , , , and , “The Effects of Entry on Incumbent Innovation and

Productivity,” Review of Economics and Statistics, 2009, 91 (1), 20–32.

Arrow, Kenneth J., “Economic Welfare and the Allocation of Resouces for Invention,”

1962. In The Rate and Direction of Inventive Activity: Economic and Social Factors,

ed. J. Neyman, 507-532. Berkeley: University of California Press.

Baldwin, William L. and John T. Scott, “Market Structure and Technological Change,”

1987. In Lesourne and Sonnenschein (Eds.), Fundamentals of Pure and Applied Eco-

nomics. Chur, Switzerland, and London: Harwook Academic Publishers.

Berry, Steven, “Estimation of a Model of Entry in the Airline Industry,” Econometrica,

1992, 60, 889–917.

Blundell, Richard, Rachel Griffith, and John Van Reenen, “Market share, market

value and innovation in a panel of British manufacturing firms,” Review of Economic

Studies, 1999, 66, 529–554.

Bresnahan, Timothy F. and Peter C. Reiss, “Entry and Competition in Concentrated

Industries,” Journal of Political Economy, 1991, 99 (5), 977–1009.

Busse, Meghan R., Christopher R. Knittel, and Florian Zettelmeyer, “Are Con-

sumers Myopic? Evidence from New and Used Car Purchases,” American Economic

Review, 2013, 103 (1), 220–256.

Ciliberto, Federico and Elie Tamer, “Market Structure and Multiple Equilibria in Air-

line Markets,” Econometrica, 2009, 77, 1791–1828.

24

Cockburn, Iain and Rebeccca Henderson, “Racing to invest? The dynamics of com-

petition in ethical drug discovery,” Journal of Economics and Management Strategy,

1994, 3, 481–519.

Cohen, Wesley M., “Empirical Studies of Innovative Activity,” 1995. In Stoneman, P.

(Ed.), Handbook of the Economics of Innovation and Technical Change. Oxford: Basil

Blackwell.

, “Fifty Years of Empirical Studies of Innovative Activity and Performance,” 2010. In

Arrow, K.J. and Intriligator, M.D. (eds), Handbook of the Economics of Innovation,

Volume 01. Elsevier.

and Richard C. Levin, “Empirical Studies of Innovation and Market Structure,”

1989. In Schmalensee, R. and Willig, R. (Eds.), Handbook of Industrial Organization.

Amsterdam: North-Holland.

Dasgupta, Partha and Joseph E. Stiglitz, “Industrial Structure and the Nature of

Innovative Activity,” The Economic Journal, 1980, 90 (358), 266–293.

Fowlie, Meredith, Mar Reguant, and Stephen P. Ryan, “Market-Based Emissions

Regulation and Industry Dynamics,” Journal of Political Economy, 2015. Forthcoming.

Gilbert, Richard, “Looking for Mr. Schumpeter: Where Are We in the Competition-

Innovation Debate?,” 2006. In Jaffe, A.B., Lerner, J. and Stern, S. (Eds.), Innovation

Policy and the Economy. Boston: The MIT Press, 159-215.

and David Newbery, “Preemptive patenting and the persistence of monopoly,”

American Economic Review, 1982, 72, 514–526.

Goettler, Ronald and Brett Gordon, “Does AMD Spur Intel to Innovate More?,” Jour-

nal of Political Economy, 2011, 119 (6), 1141–1200. Mimeo.

Grabowski, H.G. and N.D. Baxter, “Rivalry in industrial research and development:

An empirical study,” Journal of Industrial Economics, 1973, 21, 209–235.

Jaffe, Sonia and E. Glen Weyl, “The First Order Approach to Merger Analysis,” Amer-

ican Economic Journal: Microeconomics, 2013, 5 (4), 188–218.

Kamien, Morton I. and Nancy L. Schwartz, Market Structure and Innovation, Cam-

bridge: Cambridge University Press, 1982.

25

Knittel, Christopher R., “Automobile on Steroids: Product Attribute Trade-Offs and

Technological Progress in the Automobile Sector,” American Economic Review, 2011,

101 (7), 3368–3399.

Kutner, Michael H., Chris J. Nachtsheim, and John Neter, Applied Linear Regres-

sion Models, 4th ed., New York: McGraw-Hill Irwin, 2004.

Langer, Ashley and Nathan H. Miller, “Automakers’ Short-Run Responses to Changing

Gasoline Prices,” Review of Economics and Statistics, 2013, 95 (4), 1198–1211.

Lerner, Josh, “An empirical exploration of a technology race,” RAND Journal of Eco-

nomics, 1997, 28, 228–247.

Levin, Richard C., Wesley M. Cohen, and David C. Mowery, “R&D Appropri-

ability, Opportunity, and Market Structure: New Evidence on Some Schumpeterian

Hypotheses,” American Economic Review, 1985, 75 (2), 20–24.

Linn, Joshua and Thomas Klier, “New Vehicle Characteristics and the Cost of the

Corporate Average Fuel Economy Standard,” RAND Journal of Economics, 2012, 43

(1), 186–213.

Miller, Nathan H. and Matthew Osborne, “Spatial Differentiation and Price Discrim-

ination in the Cement Industry: Evidence from a Structural Model,” RAND Journal

of Economics, 2014, 45 (2), 221–247.

, Marc Remer, Conor Ryan, and Gloria Sheu, “Pass-Through and the Prediction

of Merger Price Effects,” Journal of Industrial Economics, 2015.

, Matthew Osborne, and Gloria Sheu, “Pass-Through in a Concentrated Industry:

Empirical Evidence and Regulatory Implications,” 2015.

Nickell, Stephen J., “Competition and Corporate Performance,” Journal of Political Econ-

omy, 1996, 104, 724–746.

Perez-Saiz, Hector, “Building New Plant or Entering by Acquisition? Firm Heterogeneity

and Entry Barriers in the U.S. Cement Industry,” RAND Journal of Economics, 2015,

46 (3), 625–649.

Ryan, Stephen, “The Costs of Environmental Regulation in a Concentrated Industry,”

Econometrica, 2012, 80 (3), 1019–1062.

26

Schumpeter, Joseph A., Capitalism, Socialism, and Democracy, Harper, 1942.

Scott, John T., “Firm versus Industry Variability in R&D Intensity,” 1984. In Griliches,

Zvi (Ed.), R&D, Patents, and Productivity. Chicago: University of Chicago Press, 233-

240.

Stiglitz, Joseph E. and Bruce C. Greenwald, Creating a Learning Society, Columbia

University Press, 2014.

Toivanen, Otto and Michael Waterson, “Market Structure and Entry: Where’s the

beef?,” RAND Journal of Economics, 2005, 36 (3), 680–699.

Van Oss, Hendrik G., “Background Facts and Issues Concerning Cement and Cement

Data,” 2005. Open-File Report 2005-1152, U.S. Department of the Interior, U.S. Geo-

logical Survey.

27

Appendix Materials

A Measuring Fuel Costs

We calculate the fuel costs of each kiln as the price of the primary fuel (dollars per mBtu)

multiplied by the energy requirements of production (mBtu per metric tonne of cement).

Figure B.2 shows the fraction of industry capacity that uses bituminuous coal, petroleum

coke, and natural gas as a primary fossil fuel (panel A), as well as the price of those fuels

(panel B). In the early years of the sample, bituminous coal and natural gas are the most

common primary fuel. Due a change in relative prices, however, natural gas quickly phases

out and is replaced by bituminous coal and petroleum coke.

We ascertain the energy requirements from the labor-energy input surveys of the PCA.

Conditional on kiln technology, there is no discernible change in energy efficiency over 1990-

2010, and we calculate the energy requirements of production to be 4.13, 4.32, 5.54, and 6.37

mBtu per metric tonne of cement for precalciner kilns, preheater kilns, long dry kilns, and

wet kilns, respectively, during that period. Our calculations are corroborated by a recent

USGS survey of cement plants (Van Oss (2005)). Efficiency improvements are evident over

1974-1990 within kiln type, and we assume that these are realized linearly.

Despite the transparency with which fossil fuel prices affect fuels costs, some difficulties

arise in calculating fuel costs. We highlight two here. First, plants on occasion list multiple

primary fuels in the PCA Plant Information Survey, and we use a simple decision rule in

those instances. We calculate fuel costs with the coal price of coal is listed. Otherwise we use

coke prices if coke is among the primary fuels. Otherwise we use natural gas prices if natural

gas is listed. The exception to the above decision rule is when plants use a mix of coal and

petroleum coke – there we assign equal weights to coal and petroleum coke prices. This

decision rule reflects the relative rates at which these fuels are used (e.g., see Figure B.2).

In related research, we experiment with more sophisticated allocations and the underlying

empirical relationships are unaffected (Miller, Osborne and Sheu (2015b)).

Second, our methodology does not incorporate secondary fuels, the most popular of

which are waste fuels such as solvents and used tires. We do not have data on the prices

of waste fuels, but understand them to be lower on a per-mBtu basis than fossil fuels. The

labor-energy surveys indicate that waste fuels account for around 25% of the energy used in

wet kilns and 5% in dry kilns. Our results are robust to the use of a fuel cost measure in

which fossil fuel requirements are scaled down in accordance with the survey data.

28

B Appendix Figures and Tables

29

Table B.1: Control Variables for the Main Logit Regressions

(i) (ii) (iii) (iv) (v)

Demand 0.018 0.027** 0.016 0.009(0.013) (0.012) (0.013) (0.012)

Kiln Age 0.028*** 0.032*** 0.027*** 0.027***(0.007) (0.006) (0.007) (0.007)

Kiln Capacity -0.421 0.074 -0.498 -0.572(0.784) (0.701) (0.796) (0.811)

Wet Kiln 0.367 0.351 0.351 0.329(0.377) (0.373) (0.383) (0.389)

Long Dry Kiln 0.698* 0.564 0.598* 0.526(0.361) (0.357) (0.363) (0.371)

Time Trend 0.019 0.040*** 0.019 0.024*(0.013) (0.011) (0.013) (0.013)

Notes: The regression sample includes 7,540 kiln-year observations over 1974-2008. Thedependent variable is an indicator that equals one if during the year that the kiln is replacedfor precalciner technology. The technology fixed effects include indicators for wet and longdry kilns. The standard errors, shown in parentheses, are clustered at the kiln level toaccount for autocorrelation and heteroskedastity. Statistical significance at the 10%, 5%,and 1% levels are denoted with *, **, and ***, respectively.

Table B.2: Precalciner Kilns Constructed 1974-1981

Year Firm City State Rival Capacity Mean Rival Capacity

1974 Southwestern Fairborn OH 24.58 12.361978 Lehigh Mason City IA 10.83 11.731979 Kaiser San Antonio TX 7.91 11.761979 Ideal Basic Knoxville TN 20.96 11.761980 Texas Cement New Braunfels TX 8.16 11.061980 Medusa Charlevoix MI 8.08 11.061980 Martin Marietta Lyons CO 3.25 11.061980 General New Braunfels TX 8.23 11.061981 Genstar Redding CA 3.61 11.361981 Marquette Cape Girardeau MO 20.79 11.361981 Lonestar Santa Cruz CA 9.67 11.361981 California Cement Mojave CA 8.12 11.361981 Martin Marietta Leamington UT 3.28 11.361981 Martin Marietta Buffalo IA 19.17 11.361981 Ideal Basic Theodore AL 6.57 11.361981 Alamo San Antonia TX 8.97 11.361981 Kaiser Permanente CA 8.09 11.36

Notes: Rival Capacity is the sum of capacity among competing plants within 400 miles, in millions of metrictonnes. The mean shown is the national average of Rival Capacity during the first year of the kiln’s operation.

30

0

-1

-2

-3

-4

Con

trib

utio

n

0 5 10 15 20 25Rival Capacity within 400 Miles

Figure B.1: Nonlinearities in the Effect of Competition

0

.2

.4

.6

.8

1

Per

cent

age

of C

apac

ity

1975 1980 1985 1990 1995 2000 2005 2010

Coal Gas

Coke Coal & Coke

Panel A: Capacity by Primary Fuel

0

2

4

6

8

10

Rea

l 201

0 D

olla

rs p

er m

Btu

1975 1980 1985 1990 1995 2000 2005 2010

Coal Gas

Coke

Panel B: Average Fuel Prices

Figure B.2: Primary Fuels and Fuel PricesNotes: Panel A plots the fraction of kiln capacity that burns as its primary fuel (i) bituminous coal, (ii) natural gas, (iii)petroleum coke, and (iv) bituminous coal and petroleum coke. Data are obtained from the PCA Plant Information Surveys.Panel B plots the average national prices for these fuel in real 2010 dollars per mBtu. Coal prices are obtained from the CoalReports of the Energy Information Agency (EIA); other prices are obtained from the State Energy Data System of the EIA.

31